CN106570464A - Human face recognition method and device for quickly processing human face shading - Google Patents
Human face recognition method and device for quickly processing human face shading Download PDFInfo
- Publication number
- CN106570464A CN106570464A CN201610930157.3A CN201610930157A CN106570464A CN 106570464 A CN106570464 A CN 106570464A CN 201610930157 A CN201610930157 A CN 201610930157A CN 106570464 A CN106570464 A CN 106570464A
- Authority
- CN
- China
- Prior art keywords
- class
- face
- face sample
- sample
- test
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/168—Feature extraction; Face representation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/28—Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/172—Classification, e.g. identification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/513—Sparse representations
Abstract
The invention discloses a human face recognition method and device for quickly processing human face shading, and the method comprises the steps: obtaining a test human face sample and a standardized learning dictionary; solving a local restriction code of an unshaded part of the test human face sample through employing a target function with the local restriction, and modeling a shaded region of the test human face sample through a Markov random field till the detection is completed; obtaining the sparse representation of the unshaded part of the test human face sample through the target function with l2-normal constraint; generating a reconfigured human face, corresponding to each class, class by class, and solving a reconfiguration error between the reconfigured human face of each class and the test human face sample; searching a corresponding class with the minimum reconfiguration error, determining the class as the class of the test human face sample, and outputting the class. The method can process and recognize a human face which is shaded or polluted by noise, is higher in recognition speed under the condition of guaranteeing the recognition success rate, achieves the better recognition of the human face, and is more suitable for an actual scene.
Description
Technical field
The present invention relates to computer vision and mode identification technology, and in particular to what a kind of quick process face was blocked
Face identification method and device.
Background technology
For at present, recognition of face is one of hot research problem of computer vision and area of pattern recognition, and is blocked
Problem becomes an emphasis and the difficulties in field of face identification due to its multiformity.
John Wright et al. proposed rarefaction representation classifier methods (Sparse in 2009
Representation-based classification, abbreviation SRC), this method thinks:Each test facial image
Can be represented by the linear combination of the facial image of training sample concentration identical category.Therefore, ideally, other
The corresponding code coefficient of training sample of classification is zero, and coding corresponding with test face sample class identical training sample
Coefficient is not zero, and this is embodied, and one kind is openness, and this code coefficient alternatively referred to as tests the rarefaction representation of facial image.
SRC concrete step is:(1) using the training sample image of all categories as study dictionary (2) by solving l1Minimum is asked
Topic obtains sparse coefficient (3) of the test facial image on whole study dictionary and utilizes per the corresponding sparse coefficient of class by class generation
Such reconstructed image, and reconstructed error (4) class of the selection with minimal reconstruction error of every class is obtained as final result,
Complete the identification for testing facial image.
However, when occlusion issue is solved, SRC needs to introduce a unit matrix to build a new dictionary,
Then sparse coefficient of the test face sample on new dictionary is obtained.Do so and there are problems that two:(1) new dictionary columns is very
Greatly, it is desirable to which sparse coefficient quite time-consuming (2) this way of the solution test facial image on this new dictionary can not be solved well
Certainly continuous occlusion issue.
Subsequently Zihan Zhou et al. propose the method (Sparse continuously blocked based on the process of Markov random field
Error Correction with MRF, abbreviation SEC_MRF), compared to SRC, SEC_MRF is in process sound pollution and continuously
Block aspect and all achieve good effect.However, the computational complexity of SEC_MRF is very big, it is needed to each class sample all
Iteratively solve l1Norm minimum problem, therefore arithmetic speed is slow, is not suitable for real-time scene.
The content of the invention
First shortcoming for aiming to overcome that prior art of the present invention and deficiency, there is provided a kind of quick face that processes hides
The face identification method of gear, the face identification method identification tape block or by sound pollution face when, ensureing high identification
There is recognition speed faster on the premise of success rate, so as to preferably carry out recognition of face, be more applicable for actual scene.
Further object is that overcoming the shortcoming and deficiency of prior art, there is provided a kind of quick face that processes hides
The face identification device of gear.
First purpose of the present invention is achieved through the following technical solutions:
A kind of face identification method that quick process face is blocked, the face identification method comprises the following steps:
Obtain test face sample and standard chemical handwriting practicing allusion quotation;
The local restriction for obtaining non-shield portions in the test face sample using the object function with local restriction is compiled
Code, using Markov random field is to test face sample occlusion area modeling and iteration updates, until finally having detected
Into;
Using band l2The object function of norm constraint obtains the sparse expression of the non-occluded area of the test face sample;
Generate per the corresponding reconstruct face of class by class, obtain the weight between every class reconstruct face and the test face sample
Structure error;
Search and there is the corresponding class of minimal reconstruction error, be defined as the test classification of face sample and its is defeated
Go out.
Further, the acquisition process of the standard chemical handwriting practicing allusion quotation, specifically includes:
Set up training face sample set;
Complete study dictionary was constructed using the training face sample set;
By each training sample standardization in the excessively complete study dictionary;
The known class of excessively complete study dictionary and each training sample image after by the standardization is stored.
Further, it is described to obtain the non-shield portions of the test face sample using the object function with local restriction
Local restriction is encoded, using Markov random field is to test face sample occlusion area modeling and iteration updates, until
Final detection is completed, and is specifically included:
Assume that label vector isWhether each component identification respective pixel of s is the pixel of occlusion area,
Wherein s [j]=0 represents that j elements are occlusion area pixels, and s [j]=1 represents that j elements are non-occluded area pixels, and s's is all
Element is initialized as 1;
According to current region labeling, by the way that following formula is by the non-occluded area in standard chemical handwriting practicing allusion quotation and tests face
The non-occluded area of sample image is extracted:
A*=A [st-1=1,:],y*=y [st-1=1]
In formula, A*Represent the non-occluded area of standard chemical handwriting practicing allusion quotation, y*The non-occluded area of test face sample is represented,
st-1=1 expression takes the pixel that current label vector intermediate value is 1, that is, take the non-occluded area for detecting at present;
Calculated using following formula between the non-occluded area and the non-occluded area of standard chemical handwriting practicing allusion quotation of test face sample
Similarity, and generate apart from dictionary
In formula, Ai *Represent the non-occluded area of the sub- dictionary that the i-th class training sample image is constituted, y*Represent test face sample
This non-occluded area,Y is sought in expression*And A*Between Euclidean distance, σ is one
Constant, for the control weight rate of decay;
Local restriction is applied to code coefficient, tester is obtained by the object function with local restriction in solution following formula
The code coefficient of the non-shield portions of face sample, the object function with local restriction is as follows:
In formula, | | | |2Represent the l of vector2Norm, A*Represent the non-occluded area of standard chemical handwriting practicing allusion quotation, y*Represent and survey
The non-occluded area of examination face sample, x presentation code coefficients, D represents that apart from dictionary λ is the positive number of two before and after a balance;
Code coefficient and standard chemical handwriting practicing allusion quotation according to trying to achieve is reconstructed, obtain correspondence reconstructed image and between weight
Structure error:
In formula, A represents standard chemical handwriting practicing allusion quotation, and y is represented and tested face sample, x presentation code coefficients,Represent that reconstruct is missed
Difference;
Markov random field model is set up to label vector s, the object function in Algorithm for Solving following formula is cut using figure, more
New label vector s:
Wherein,Value as described in following formula:
In formula, s [i] is the label value of ith pixel,It is the reconstructed error value of ith pixel, μ is a constant,
The intensity of the interphase interaction of control different pixels, τ is a threshold value;
Iteration performs above-mentioned steps, until algorithmic statement or reaching maximum convergence number of times, now tests face sample reasonable
Band occlusion area be detected.
Further, the object function with local restriction has analytic solutions, and its solution procedure is specifically included:
Obtain data covariance matrix:
S=(A*T-1y*T)(A*T-1y*T)T
In formula, S represents covariance matrix, A*Represent the non-occluded area of standard chemical handwriting practicing allusion quotation, y*Represent test face sample
This non-occluded area, 1 vector representation element value is all 1 vector, it is assumed that non-occluded area pixel quantity is p, then
Obtain code coefficient x:
X=(S+ λ diag (D)2)\1
In formula, S represents covariance matrix, and λ is the positive number of two before and after balance in above formula object function, and diag (D) is right
Diagonal element is the diagonal matrix of element in D, and 1 vector representation element value is all 1 vector, a the inverse of b representing matrix a be multiplied by square
Battle array b;
Code coefficient x is normalized, normalized code coefficient is obtained
In formula, x presentation code coefficients, 1 vector representation element value is all 1 vector, and a/b representing matrix a are multiplied by matrix b's
It is inverse.
Further, the object function with local restriction can be simplified to following formula:
In formula, B=A*(:, id), id is the minimum corresponding index of element in dictionary, A*Represent standard chemical handwriting practicing
The non-occluded area of allusion quotation.
Further, it is described using band l2The object function of norm constraint obtains the unshielede region of the test face sample
The sparse expression in domain, specially:
Using solution l2The method of constrained minimization, by band l in solution following formula2The object function of norm constraint is to the survey
The non-occluded area part of examination face sample carries out l2Sparse coding:
In formula, | | | |2Represent the l of vector2Norm, x represents sparse coding coefficient, argminx||x||2Expression takes makes x
L2The corresponding x of Norm minimum, y*Represent the non-occluded area of test face sample, A*Represent not hiding for standard chemical handwriting practicing allusion quotation
Gear region, θ is the positive number of two before and after a balance.
Further, the band l2The object function of norm constraint has analytic solutions, and it was solved
Journey is specifically included:Obtain projection matrix P:
In formula, A*The non-occluded area of standard chemical handwriting practicing allusion quotation is represented, I matrixes are unit matrixs, ()-1Expression takes inverse behaviour
Make, θ is the positive number of two before and after balance in object function described in S109;
Obtain l2Sparse coding:
In formula,Represent l2Sparse coding coefficient, y*Represent the non-occluded area of test face sample.
Further, it is described to generate per the corresponding reconstruct face of class by class, obtain every class reconstruct face and the tester
Reconstructed error between face sample, specially:
Using the l for solving2Sparse coding coefficient, by class such corresponding reconstruct face and reconstruct face and original are obtained
Reconstructed error between test face sample:
In formula,Represent the reconstructed error of the i-th class, y*Represent the non-occluded area of test face sample, Ai *Represent the i-th class
The non-occluded area of the sub- dictionary of training sample image composition,Represent the sparse coding coefficient of the i-th class.
Further, described lookup has the corresponding class of minimal reconstruction error, is defined as the test face sample
Classification and output it, specially:
According to the reconstructed error of each class, the class with minimal reconstruction error is searched as the class of the test face sample
Not:
In formula, identity (y) represents the classification of test face sample,The reconstructed error of the i-th class is represented,Represent the
The sparse coding coefficient of i classes,Expression takes to be madeMinimum classification i.
Another object of the present invention is achieved through the following technical solutions:
A kind of face identification device that quick process face is blocked, the face identification device includes:
Acquisition module, for obtaining test face sample and standard chemical handwriting practicing allusion quotation;
Detection module, for obtaining the non-shield portions of the test face sample using the object function with local restriction
Local restriction is encoded, using Markov random field is to test face sample occlusion area modeling and iteration updates, until
Final detection is completed;
Sparse coding module, for using band l2The object function of norm constraint obtains not hiding for the test face sample
The sparse expression in gear region;
Reconstructed error module is calculated, for using generating per the corresponding reconstruct face of class by class, obtaining every class reconstruct face
With the reconstructed error between the test face sample;
Output module, has the corresponding class of minimal reconstruction error for searching, and is defined as the test face sample
Classification and output it.
Wherein, the acquisition module includes:
Test sample submodule is obtained, for obtaining face image data and being converted into test face sample;
Training sample set submodule is set up, for setting up training face sample set;
Study dictionary submodule is obtained, for constructing complete study dictionary using the training face sample set;
Standard chemical handwriting practicing allusion quotation submodule, for by described study dictionary each training sample standardization;
Storage standard chemical handwriting practicing allusion quotation submodule, for by the study dictionary and each training sample after the standardization
The known class of image is stored.
Wherein, the detection module includes:
Non-occluded area submodule is obtained, before obtaining the test face sample and standard chemical handwriting practicing mortgage
With the pixel for not blocking label;
Generate apart from dictionary submodule, non-occluded area is currently demarcated with the mark for calculating the test face sample
Euclidean distance in quasi- chemistry handwriting practicing allusion quotation between the current demarcation non-occluded area of each training sample image, and generate distance
Dictionary;
Local restriction encoding submodule, for calculating the survey apart from dictionary, the standard chemical handwriting practicing allusion quotation using described
The local restriction coding of examination face sample non-occluded area;
Calculation code error submodule, for calculating the local restriction coding life of the test face sample non-occluded area
Encoding error between reconstruct face and the test face sample;
Occlusion area submodule is updated, the test face sample occlusion area is modeled simultaneously using Markov random field
Occlusion area is updated according to the encoding error;
Judging submodule, for judging whether to reach maximum iteration time, if it is, into the sparse volume of non-occluded area
Code module, conversely, then returning to acquisition non-occluded area sub-blocks of pixels.
Compared with prior art, the beneficial effects of the present invention is:
1st, the present invention is carrying out drawing in detection process using Markov random field to the occlusion area for testing face sample
Enter local bound term, the code coefficient of non-occluded area is obtained using local restriction item, both considered mutual between pixel
Effect, the dissimilarity between similarity and face between face being efficiently utilized again and being blocked is effectively guaranteed
High discrimination.
2nd, while ensureing high discrimination, during iteration updates occlusion area, because local restriction item can derive solution
Analysis solution, effectively reduces computational complexity, so as to accelerate the process of detection occlusion area.Meanwhile, after occlusion area has been detected,
L is used to non-occluded area2Sparse coding completes identification process, l2Minimization problem can equally derive analytic solutions, equally reduce
The time required to computing.By such mode, the present invention is effectively improved recognition speed, reaches what quick process face was blocked
Purpose.
Description of the drawings
Fig. 1 is a kind of stream of specific embodiment of the face identification method that quick process face disclosed by the invention is blocked
Journey schematic diagram;
Fig. 2 is that the face identification method that quick process face disclosed by the invention is blocked is smart with the classification of other five kinds of methods
Spend as random shielded area becomes curve chart that is big and changing;
Fig. 3 is a kind of knot of specific embodiment of the face identification device that quick process face disclosed by the invention is blocked
Structure block diagram.
Specific embodiment
In order that those skilled in the art more fully understand the present invention program, below in conjunction with accompanying drawing and specific embodiment party
Formula, is described in further detail to the present invention, it is clear that described embodiment is only the embodiment of a present invention part, and
It is not all, of embodiment.Based on the embodiment in the present invention, those of ordinary skill in the art are not making creative work
Under the premise of the every other embodiment that obtained, belong to the scope of protection of the invention.
Embodiment one
Present embodiment discloses the face identification method that a kind of quick process face is blocked, it is therefore intended that block in identification tape
Or by sound pollution face when, have recognition speed faster on the premise of high recognition success rate is ensured, so as to preferably enter
Row recognition of face.In the present embodiment the face identification method that a kind of quick process face is blocked schematic flow sheet as shown in figure 1,
Specifically include following steps:
S1, acquisition test face sample and standard chemical handwriting practicing allusion quotation.The sample image for assuming test is y, dimension and instruction
Practice sample image consistent, be equally also converted into the column vector of m × 1, thenAssuming the sample image of training has a classes,
Have b sample images per class, then the sample image of training is a total of n=a × b, the length of every image and it is wide be respectively f and
G, then the dimension of every image is m=f × g, and every training sample image is converted to into the column vector of m × 1 by f × g, and is generated
One standard chemical handwriting practicing allusion quotation
Standard chemical handwriting practicing allusion quotation in step S1 pre-builds, pre-builds the process of standard chemical handwriting practicing allusion quotation
Including:
S101, foundation training face sample set;
S102, using it is described training face sample set constructed complete study dictionary;
S103, by described study dictionary each training sample standardization;
S104, the study dictionary and the known class of each training sample image after the standardization is stored.
S2, hypothesis label vector areWhether each component identification respective pixel of s is the picture of occlusion area
Element, wherein s [j]=0 represent that j elements are occlusion area pixels, and all elements of s [j]=1, s are initialized as 1.
S3, according to current region labeling, by following formula by the non-occluded area in standard chemical handwriting practicing allusion quotation and test
The non-occluded area of face sample is extracted:
A*=A [st-1=1,:],y*=y [st-1=1]
In formula, A*Represent the non-occluded area of standard chemical handwriting practicing allusion quotation, y*The non-occluded area of test face sample is represented,
st-1=1 expression takes the pixel that current label vector intermediate value is 1, that is, take the non-occluded area for detecting at present.
S4, using following formula calculate test face sample non-occluded area and standard chemical handwriting practicing allusion quotation non-occluded area it
Between similarity, and generate apart from dictionary
In formula, Ai *Represent the non-occluded area of the sub- dictionary that the i-th class training sample image is constituted, y*Represent test face sample
This non-occluded area,Y is sought in expression*And A*Between Euclidean distance, σ is one
Constant, for the control weight rate of decay.
S5, local restriction is applied to code coefficient, by solving following formula in object function testing face sample not
The code coefficient of shield portions:
In formula, | | | |2Represent the l of vector2Norm, A*Represent the non-occluded area of standard chemical handwriting practicing allusion quotation, y*Represent and survey
The non-occluded area of examination face sample, x presentation code coefficients, D represents that apart from dictionary λ is the positive number of two before and after a balance;
The solution of the object function of the non-shield portions code coefficient of test face sample in described step S5 has solution
Analysis is solved, described in following surface current journey:
S501, obtain data covariance matrix:
S=(A*T-1y*T)(A*T-1y*T)T
In formula, S represents covariance matrix, A*Represent the non-occluded area of standard chemical handwriting practicing allusion quotation, y*Represent test face sample
This non-occluded area, 1 vector representation element value is all 1 vector.It is assumed here that non-occluded area pixel quantity is p, then
S502, obtain code coefficient x:
X=(S+ λ diag (D)2)\1
In formula, S represents covariance matrix, and λ is the positive number of two before and after balance in object function described in S05, and diag (D) is
Diagonal entry is the diagonal matrix of element in D, and 1 vector representation element value is all 1 vector, a the inverse of b representing matrix a be multiplied by
Matrix b;
S503, code coefficient x is normalized, obtains normalized code coefficient
In formula, x presentation code coefficients, 1 vector representation element value is all 1 vector, and a/b representing matrix a are multiplied by matrix b's
It is inverse;
Meanwhile, the object function of the non-shield portions code coefficient of test face sample in step S5 can be further
It is simplified to following formula:
In formula, B=A*(:, id), id is the minimum corresponding index of element in dictionary, A*Represent standard chemical handwriting practicing
The non-occluded area of allusion quotation.By so simplifying, the computational complexity for solving code coefficient is again reduced, arithmetic speed is entered
The lifting of one step.
S6, be reconstructed according to the code coefficient obtained and study dictionary, obtain correspondence reconstructed image and between reconstruct
Error:
In formula, A represents standard chemical handwriting practicing allusion quotation, and y is represented and tested face sample, x presentation code coefficients,Represent that reconstruct is missed
Difference.
S7, Markov random field model is set up to label vector s, the target letter in Algorithm for Solving following formula is cut using figure
Number, updates label vector s:
Wherein,Value as described in following formula:
In formula, s [i] is the label value of ith pixel,It is the reconstructed error value of ith pixel, μ is a constant,
The intensity of the interphase interaction of control different pixels, τ is a threshold value.
S8, iteration perform S3-S7, until algorithmic statement or reaching maximum convergence number of times, now test face sample reasonable
Band occlusion area be detected.
S9, employing solution l2The method of constrained minimization, by the object function in solution following formula to testing face sample y's
Non-occluded area part carries out l2Sparse coding:
In formula, | | | |2Represent the l of vector2Norm, x represents sparse coding coefficient, argminx||x||2Expression takes makes x
L2The corresponding x of Norm minimum, y*Represent the non-occluded area of test face sample, A*Represent not hiding for standard chemical handwriting practicing allusion quotation
Gear region, θ is the positive number of two before and after a balance;
Test face sample non-occluded area l in step S92The solution of the object function of sparse coding has solution
Analysis is solved, described in following surface current journey:
S901, obtain projection matrix P:
In formula, A*The non-occluded area of standard chemical handwriting practicing allusion quotation is represented, I matrixes are unit matrixs, ()-1Expression takes inverse behaviour
Make, θ is the positive number of two before and after balance in object function described in S109;
S902, obtain l2Sparse coding:
In formula,Represent l2Sparse coding coefficient, y*Represent the non-occluded area of test face sample.
The sparse coding that S10, utilization are solved, such corresponding reconstructed image and the reconstructed image are obtained with original by class
Reconstructed error between test image:
In formula,Represent the reconstructed error of the i-th class, y*Represent the non-occluded area of test face sample, Ai *Represent the i-th class
The non-occluded area of the sub- dictionary of training sample image composition,Represent the sparse coding coefficient of the i-th class.
S11, according to the reconstructed error of each class, select the class with minimal reconstruction error as the classification of y:
In formula, identity (y) represents test face sample class,The reconstructed error of the i-th class is represented,Represent i-th
The sparse coding coefficient of class,Expression takes to be madeMinimum classification i;
The recognition result of S12, output test face sample.
Fig. 2 gives the nicety of grading of six kinds of algorithms as random shielded area becomes curve chart that is big and changing, and table 1 is given
Six kinds of algorithms process 50% block at random when run times.Six kinds of control methods are respectively:SRC, SEC_MRF,
Collaborative Representation based Classification with Regularized Least
Square (abbreviation CRC_RLS), Robust Sparse Coding (abbreviation RSC), Sparse Illumination
Transfer (abbreviation SIT) and the present invention.Even if can see in the case of 50% shielded area, the discrimination of the present invention
Only more lower slightly than SEC_MRF 0.7%, but the speed of service is but 2.5 times or so of SEC_MRF, therefore it is more suitable for real-time scene.
Run time of the 1. 6 kinds of algorithms of table when process 50% is blocked at random
Embodiment two
Present invention also offers the face identification device that a kind of quick process face is blocked, a kind of knot of specific embodiment
Structure schematic diagram is as shown in figure 3, the device includes:
Acquisition module 100, for obtaining test face sample and standard chemical handwriting practicing allusion quotation;
Detection module 200, for obtaining the non-occlusion part of the test face sample using the object function with local restriction
The local restriction coding for dividing, using Markov random field to test face sample occlusion area modeling simultaneously iteration renewal,
Until final detection is completed;
Sparse coding module 300, for using band l2The object function of norm constraint obtains the test face sample
The sparse expression of non-occluded area;
Reconstructed error module 400 is calculated, for using generating per the corresponding reconstruct face of class by class, obtaining every class reconstruct people
Reconstructed error between face and the test face sample;
Output module 500, has the corresponding class of minimal reconstruction error for searching, and is defined as the test face sample
This classification simultaneously outputs it.
In concrete application, the acquisition module 100 includes:
Test sample submodule 101 is obtained, for obtaining face image data and being converted into test face sample;
Training sample set submodule 102 is set up, for setting up training face sample set;
Study dictionary submodule 103 is obtained, for constructing complete study dictionary using the training face sample set;
Standard chemical handwriting practicing allusion quotation submodule 104, for by described study dictionary each training sample standardization;
Storage standard chemical handwriting practicing allusion quotation submodule 105, for by the study dictionary after the standardization and each training
The known class of sample image is stored.
In concrete application, the detection module 200 includes:
Non-occluded area submodule 201 is obtained, for obtaining the test face sample and the standard chemical handwriting practicing allusion quotation
It is current to carry the pixel for not blocking label;
Generate apart from dictionary submodule 202, for calculating the test face sample non-occluded area and institute are currently demarcated
The Euclidean distance between the current demarcation non-occluded area of each training sample image in standard chemical handwriting practicing allusion quotation is stated, and is generated
Apart from dictionary;
Local restriction encoding submodule 203, for calculating described apart from dictionary, the standard chemical handwriting practicing allusion quotation using described
The local restriction coding of test face sample non-occluded area;
Calculation code error submodule 204, the local restriction for calculating the test face sample non-occluded area is compiled
Code generates the encoding error between reconstruct face and the test face sample;
Occlusion area submodule 205 is updated, the test face sample occlusion area is built using Markov random field
Mould simultaneously updates occlusion area according to the encoding error;
Judging submodule 206, for judging whether to reach maximum iteration time, if it is, dilute into non-occluded area
Thin coding module, conversely, then returning to acquisition non-occluded area sub-blocks of pixels.
The face identification device that blocks of quick process face that the present invention is provided other concrete settings are carried with above-described embodiment
For face identification method it is similar, will not be described here.
In sum, a kind of quick process face provided by the present invention is blocked face identification method and device, first
The shield portions of test face sample are quickly recognized, then the non-shield portions to testing face sample carry out sparse coding,
Reconstructed error is calculated by class, and the classification of test face sample is determined according to kind judging formula.
It should be noted that in said apparatus embodiment, included modules simply carry out drawing according to function logic
Point, but above-mentioned division is not limited to, as long as corresponding function can be realized;In addition, the specific name of each module
Also only to facilitate mutually distinguishing, it is not limited to protection scope of the present invention.
The foregoing description of the disclosed embodiments, enables professional and technical personnel in the field to realize or using the present invention.
Various modifications to embodiment will be apparent for those skilled in the art, general original defined herein
Reason can be realized in other embodiments without departing from the spirit or scope of the present invention.Therefore, the present invention will not
It is restricted to embodiment illustrated herein, but accords with principles disclosed herein and consistent most wide of features of novelty
Scope.
Claims (10)
1. the face identification method that a kind of quick process face is blocked, it is characterised in that the face identification method includes following
Step:
Obtain test face sample and standard chemical handwriting practicing allusion quotation;
The local restriction coding of non-shield portions in the test face sample, profit are obtained using the object function with local restriction
With Markov random field is to test face sample occlusion area modeling and iteration updates, until final detection is completed;
Using band l2The object function of norm constraint obtains the sparse expression of the non-occluded area of the test face sample;
Generate per the corresponding reconstruct face of class by class, the reconstruct obtained between every class reconstruct face and the test face sample is missed
Difference;
Search and there is the corresponding class of minimal reconstruction error, be defined as the classification of the test face sample and output it.
2. the face identification method that a kind of quick process face according to claim 1 is blocked, it is characterised in that the mark
The acquisition process of quasi- chemistry handwriting practicing allusion quotation, specifically includes:
Set up training face sample set;
Complete study dictionary was constructed using the training face sample set;
By each training sample standardization in the excessively complete study dictionary;
The known class of excessively complete study dictionary and each training sample image after by the standardization is stored.
3. the face identification method that a kind of quick process face according to claim 1 is blocked, it is characterised in that the profit
The local restriction coding of the non-shield portions of the test face sample is obtained with the object function with local restriction, can using Ma Er
Husband's random field is to test face sample occlusion area modeling and iteration updates, until final detection is completed, specifically includes:
Assume that label vector isWhether each component identification respective pixel of s is the pixel of occlusion area, wherein s
[j]=0 represents that j elements are occlusion area pixels, and s [j]=1 represents that j elements are non-occluded area pixels, all elements of s
It is initialized as 1;
According to current region labeling, by the way that following formula is by the non-occluded area in standard chemical handwriting practicing allusion quotation and tests face sample
The non-occluded area of image is extracted:
A*=A [st-1=1,:],y*=y [st-1=1]
In formula, A*Represent the non-occluded area of standard chemical handwriting practicing allusion quotation, y*Represent the non-occluded area of test face sample, st-1=
1 expression takes the pixel that current label vector intermediate value is 1, that is, take the non-occluded area for detecting at present;
The phase between non-occluded area of the non-occluded area of test face sample with standard chemical handwriting practicing allusion quotation is calculated using following formula
Like degree, and generate apart from dictionary
In formula, Ai *Represent the non-occluded area of the sub- dictionary that the i-th class training sample image is constituted, y*Represent test face sample
Non-occluded area,Y is sought in expression*And A*Between Euclidean distance, σ is a constant,
For the control weight rate of decay;
Local restriction is applied to code coefficient, test face sample is obtained by the object function with local restriction in solution following formula
The code coefficient of this non-shield portions, the object function with local restriction is as follows:
In formula, | | | |2Represent the l of vector2Norm, A*Represent the non-occluded area of standard chemical handwriting practicing allusion quotation, y*Represent tester
The non-occluded area of face sample, x presentation code coefficients, D represents that apart from dictionary λ is the positive number of two before and after a balance;
Code coefficient and standard chemical handwriting practicing allusion quotation according to trying to achieve is reconstructed, obtain correspondence reconstructed image and between reconstruct miss
Difference:
In formula, A represents standard chemical handwriting practicing allusion quotation, and y is represented and tested face sample, x presentation code coefficients,Represent reconstructed error;
Markov random field model is set up to label vector s, using figure the object function in Algorithm for Solving following formula is cut, update mark
Sign vector s:
Wherein,Value as described in following formula:
In formula, s [i] is the label value of ith pixel,It is the reconstructed error value of ith pixel, μ is a constant, control
The intensity of the interphase interaction of different pixels, τ is a threshold value;
Iteration performs above-mentioned steps, until algorithmic statement or reaching maximum convergence number of times, now tests the rational band of face sample
Occlusion area is detected.
4. the face identification method that a kind of quick process face according to claim 3 is blocked, it is characterised in that the band
The object function of local restriction has analytic solutions, and its solution procedure is specifically included:
Obtain data covariance matrix:
S=(A*T-1y*T)(A*T-1y*T)T
In formula, S represents covariance matrix, A*Represent the non-occluded area of standard chemical handwriting practicing allusion quotation, y*Represent test face sample
Non-occluded area, 1 vector representation element value is all 1 vector, it is assumed that non-occluded area pixel quantity is p, then
Obtain code coefficient x:
X=(S+ λ diag (D)2)\1
In formula, S represents covariance matrix, and λ is the positive number of two before and after balance in above formula object function, and diag (D) is diagonal
Element is the diagonal matrix of element in D, and 1 vector representation element value is all 1 vector, a the inverse of b representing matrix a be multiplied by matrix b;
Code coefficient x is normalized, normalized code coefficient is obtained
In formula, x presentation code coefficients, 1 vector representation element value is all 1 vector, and a/b representing matrix a are multiplied by the inverse of matrix b.
5. the face identification method that a kind of quick process face according to claim 3 is blocked, it is characterised in that
The object function with local restriction can be simplified to following formula:
In formula, B=A*(:, id), id is the minimum corresponding index of element in dictionary, A*Represent standard chemical handwriting practicing allusion quotation
Non-occluded area.
6. the face identification method that a kind of quick process face according to claim 1 is blocked, it is characterised in that
It is described to utilize band l2The object function of norm constraint obtains the sparse expression of the non-occluded area of the test face sample,
Specially:
Using solution l2The method of constrained minimization, by band l in solution following formula2The object function of norm constraint is to the tester
The non-occluded area part of face sample carries out l2Sparse coding:
In formula, | | | |2Represent the l of vector2Norm, x represents sparse coding coefficient, arg minx||x||2Expression takes the l for making x2
The corresponding x of Norm minimum, y*Represent the non-occluded area of test face sample, A*Represent the unshielede region of standard chemical handwriting practicing allusion quotation
Domain, θ is the positive number of two before and after a balance.
7. the face identification method that a kind of quick process face according to claim 6 is blocked, it is characterised in that
The band l2The object function of norm constraint has analytic solutions, and its solution procedure is specifically included:
Obtain projection matrix P:
P=(A*TA*+θ·I)-1A*T
In formula, A*The non-occluded area of standard chemical handwriting practicing allusion quotation is represented, I matrixes are unit matrixs, ()-1Expression takes inverse operation, θ
It is the positive number of two before and after balance in object function described in S109;
Obtain l2Sparse coding:
In formula,Represent l2Sparse coding coefficient, y*Represent the non-occluded area of test face sample.
8. the face identification method that a kind of quick process face according to claim 1 is blocked, it is characterised in that
It is described to generate per the corresponding reconstruct face of class by class, obtain the weight between every class reconstruct face and the test face sample
Structure error, specially:
Using the l for solving2Sparse coding coefficient, by class such corresponding reconstruct face and reconstruct face and former test are obtained
Reconstructed error between face sample:
In formula,Represent the reconstructed error of the i-th class, y*Represent the non-occluded area of test face sample, Ai *Represent that the i-th class is trained
The non-occluded area of the sub- dictionary of sample image composition,Represent the sparse coding coefficient of the i-th class.
9. the face identification method that a kind of quick process face according to claim 1 is blocked, it is characterised in that described to look into
Class corresponding with minimal reconstruction error is looked for, is defined as the classification of the test face sample and is output it, specially:
According to the reconstructed error of each class, the class with minimal reconstruction error is searched as the classification of the test face sample:
In formula, identity (y) represents the classification of test face sample,The reconstructed error of the i-th class is represented,Represent the i-th class
Sparse coding coefficient,Expression takes to be madeMinimum classification i.
10. the face identification device that a kind of quick process face is blocked, it is characterised in that the face identification device includes:
Acquisition module, for obtaining test face sample and standard chemical handwriting practicing allusion quotation;
Detection module, for obtaining the local of the non-shield portions of the test face sample using the object function with local restriction
Constraint coding, using Markov random field is to test face sample occlusion area modeling and iteration updates, until finally
Detection is completed;
Sparse coding module, for using band l2The object function of norm constraint obtains the unshielede region of the test face sample
The sparse expression in domain;
Reconstructed error module is calculated, for using generating per the corresponding reconstruct face of class by class, obtaining every class reconstruct face and institute
State the reconstructed error between test face sample;
Output module, has the corresponding class of minimal reconstruction error for searching, and is defined as the class of the test face sample
Not and do not output it.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610930157.3A CN106570464B (en) | 2016-10-31 | 2016-10-31 | Face recognition method and device for rapidly processing face shielding |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610930157.3A CN106570464B (en) | 2016-10-31 | 2016-10-31 | Face recognition method and device for rapidly processing face shielding |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106570464A true CN106570464A (en) | 2017-04-19 |
CN106570464B CN106570464B (en) | 2020-05-22 |
Family
ID=58534594
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610930157.3A Active CN106570464B (en) | 2016-10-31 | 2016-10-31 | Face recognition method and device for rapidly processing face shielding |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106570464B (en) |
Cited By (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107169447A (en) * | 2017-05-12 | 2017-09-15 | 贵州中信云联科技有限公司 | Hospital self-service system based on recognition of face |
CN107392211A (en) * | 2017-07-19 | 2017-11-24 | 苏州闻捷传感技术有限公司 | The well-marked target detection method of the sparse cognition of view-based access control model |
CN107729886A (en) * | 2017-11-24 | 2018-02-23 | 北京小米移动软件有限公司 | The processing method and processing device of facial image |
CN107944344A (en) * | 2017-10-30 | 2018-04-20 | 国网浙江省电力公司绍兴供电公司 | Power supply enterprise's construction mobile security supervision platform |
CN108171215A (en) * | 2018-01-25 | 2018-06-15 | 河南大学 | Face Pseudo-median filter and camouflage category detection method based on low-rank variation dictionary and rarefaction representation classification |
CN108875445A (en) * | 2017-05-08 | 2018-11-23 | 上海荆虹电子科技有限公司 | A kind of pedestrian recognition methods and device again |
CN109063506A (en) * | 2018-07-09 | 2018-12-21 | 江苏达实久信数字医疗科技有限公司 | Privacy processing method for medical operating teaching system |
CN109241852A (en) * | 2018-08-10 | 2019-01-18 | 广州杰赛科技股份有限公司 | Face identification method, device with supplementary features, computer equipment |
CN109711283A (en) * | 2018-12-10 | 2019-05-03 | 广东工业大学 | A kind of joint doubledictionary and error matrix block Expression Recognition algorithm |
CN109886167A (en) * | 2019-02-01 | 2019-06-14 | 中国科学院信息工程研究所 | One kind blocking face identification method and device |
CN110188718A (en) * | 2019-06-04 | 2019-08-30 | 南京大学 | It is a kind of based on key frame and joint sparse indicate without constraint face identification method |
CN110348331A (en) * | 2019-06-24 | 2019-10-18 | 深圳和而泰家居在线网络科技有限公司 | Face identification method and electronic equipment |
CN111079715A (en) * | 2020-01-02 | 2020-04-28 | 华南理工大学 | Occlusion robustness face alignment method based on double-dictionary learning |
CN111956203A (en) * | 2020-08-18 | 2020-11-20 | 广州视源电子科技股份有限公司 | Electrocardiosignal parameterization method, model training method, device, equipment and medium |
CN112380995A (en) * | 2020-11-16 | 2021-02-19 | 华南理工大学 | Face recognition method and system based on deep feature learning in sparse representation domain |
CN113269137A (en) * | 2021-06-18 | 2021-08-17 | 常州信息职业技术学院 | Non-fit face recognition method combining PCANet and shielding positioning |
CN114049668A (en) * | 2021-11-15 | 2022-02-15 | 北京计算机技术及应用研究所 | Face recognition method |
WO2022037541A1 (en) * | 2020-08-20 | 2022-02-24 | 腾讯科技(深圳)有限公司 | Image processing model training method and apparatus, device, and storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104200194A (en) * | 2014-08-06 | 2014-12-10 | 北京工业大学 | Face recognition method |
US9336433B1 (en) * | 2013-07-24 | 2016-05-10 | University Of Central Florida Research Foundation, Inc. | Video face recognition |
CN105740790A (en) * | 2016-01-25 | 2016-07-06 | 南京信息工程大学 | Multicore dictionary learning-based color face recognition method |
-
2016
- 2016-10-31 CN CN201610930157.3A patent/CN106570464B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9336433B1 (en) * | 2013-07-24 | 2016-05-10 | University Of Central Florida Research Foundation, Inc. | Video face recognition |
CN104200194A (en) * | 2014-08-06 | 2014-12-10 | 北京工业大学 | Face recognition method |
CN105740790A (en) * | 2016-01-25 | 2016-07-06 | 南京信息工程大学 | Multicore dictionary learning-based color face recognition method |
Non-Patent Citations (1)
Title |
---|
邓承志 等: "《基于局部约束群稀疏的红外图像超分辨重建》", 《物理学报》 * |
Cited By (28)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108875445A (en) * | 2017-05-08 | 2018-11-23 | 上海荆虹电子科技有限公司 | A kind of pedestrian recognition methods and device again |
CN108875445B (en) * | 2017-05-08 | 2020-08-25 | 深圳荆虹科技有限公司 | Pedestrian re-identification method and device |
CN107169447A (en) * | 2017-05-12 | 2017-09-15 | 贵州中信云联科技有限公司 | Hospital self-service system based on recognition of face |
CN107392211B (en) * | 2017-07-19 | 2021-01-15 | 苏州闻捷传感技术有限公司 | Salient target detection method based on visual sparse cognition |
CN107392211A (en) * | 2017-07-19 | 2017-11-24 | 苏州闻捷传感技术有限公司 | The well-marked target detection method of the sparse cognition of view-based access control model |
CN107944344A (en) * | 2017-10-30 | 2018-04-20 | 国网浙江省电力公司绍兴供电公司 | Power supply enterprise's construction mobile security supervision platform |
CN107729886A (en) * | 2017-11-24 | 2018-02-23 | 北京小米移动软件有限公司 | The processing method and processing device of facial image |
CN108171215A (en) * | 2018-01-25 | 2018-06-15 | 河南大学 | Face Pseudo-median filter and camouflage category detection method based on low-rank variation dictionary and rarefaction representation classification |
CN108171215B (en) * | 2018-01-25 | 2023-02-03 | 河南大学 | Face camouflage detection and camouflage type detection method based on low-rank variation dictionary and sparse representation classification |
CN109063506A (en) * | 2018-07-09 | 2018-12-21 | 江苏达实久信数字医疗科技有限公司 | Privacy processing method for medical operating teaching system |
CN109063506B (en) * | 2018-07-09 | 2021-07-06 | 江苏达实久信数字医疗科技有限公司 | Privacy processing method for medical operation teaching system |
CN109241852A (en) * | 2018-08-10 | 2019-01-18 | 广州杰赛科技股份有限公司 | Face identification method, device with supplementary features, computer equipment |
CN109711283B (en) * | 2018-12-10 | 2022-11-15 | 广东工业大学 | Occlusion expression recognition method combining double dictionaries and error matrix |
CN109711283A (en) * | 2018-12-10 | 2019-05-03 | 广东工业大学 | A kind of joint doubledictionary and error matrix block Expression Recognition algorithm |
CN109886167A (en) * | 2019-02-01 | 2019-06-14 | 中国科学院信息工程研究所 | One kind blocking face identification method and device |
CN110188718A (en) * | 2019-06-04 | 2019-08-30 | 南京大学 | It is a kind of based on key frame and joint sparse indicate without constraint face identification method |
CN110348331A (en) * | 2019-06-24 | 2019-10-18 | 深圳和而泰家居在线网络科技有限公司 | Face identification method and electronic equipment |
CN110348331B (en) * | 2019-06-24 | 2022-01-14 | 深圳数联天下智能科技有限公司 | Face recognition method and electronic equipment |
CN111079715A (en) * | 2020-01-02 | 2020-04-28 | 华南理工大学 | Occlusion robustness face alignment method based on double-dictionary learning |
CN111079715B (en) * | 2020-01-02 | 2023-04-25 | 华南理工大学 | Occlusion robustness face alignment method based on double dictionary learning |
CN111956203A (en) * | 2020-08-18 | 2020-11-20 | 广州视源电子科技股份有限公司 | Electrocardiosignal parameterization method, model training method, device, equipment and medium |
WO2022037541A1 (en) * | 2020-08-20 | 2022-02-24 | 腾讯科技(深圳)有限公司 | Image processing model training method and apparatus, device, and storage medium |
CN112380995A (en) * | 2020-11-16 | 2021-02-19 | 华南理工大学 | Face recognition method and system based on deep feature learning in sparse representation domain |
CN112380995B (en) * | 2020-11-16 | 2023-09-12 | 华南理工大学 | Face recognition method and system based on deep feature learning in sparse representation domain |
CN113269137A (en) * | 2021-06-18 | 2021-08-17 | 常州信息职业技术学院 | Non-fit face recognition method combining PCANet and shielding positioning |
CN113269137B (en) * | 2021-06-18 | 2023-10-31 | 常州信息职业技术学院 | Non-matching face recognition method combining PCANet and shielding positioning |
CN114049668A (en) * | 2021-11-15 | 2022-02-15 | 北京计算机技术及应用研究所 | Face recognition method |
CN114049668B (en) * | 2021-11-15 | 2024-04-09 | 北京计算机技术及应用研究所 | Face recognition method |
Also Published As
Publication number | Publication date |
---|---|
CN106570464B (en) | 2020-05-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106570464A (en) | Human face recognition method and device for quickly processing human face shading | |
CN111709409B (en) | Face living body detection method, device, equipment and medium | |
US11574152B2 (en) | Recognition system for security check and control method thereof | |
CN105740909B (en) | Text recognition method under a kind of natural scene based on spatial alternation | |
CN109359559B (en) | Pedestrian re-identification method based on dynamic shielding sample | |
CN112966684A (en) | Cooperative learning character recognition method under attention mechanism | |
CN111950453A (en) | Optional-shape text recognition method based on selective attention mechanism | |
CN107590432A (en) | A kind of gesture identification method based on circulating three-dimensional convolutional neural networks | |
CN113010656B (en) | Visual question-answering method based on multi-mode fusion and structural control | |
CN112419327A (en) | Image segmentation method, system and device based on generation countermeasure network | |
CN111832484A (en) | Loop detection method based on convolution perception hash algorithm | |
CN113344206A (en) | Knowledge distillation method, device and equipment integrating channel and relation feature learning | |
CN110956158A (en) | Pedestrian shielding re-identification method based on teacher and student learning frame | |
Li et al. | Image manipulation localization using attentional cross-domain CNN features | |
CN112418032A (en) | Human behavior recognition method and device, electronic equipment and storage medium | |
CN115240259A (en) | Face detection method and face detection system based on YOLO deep network in classroom environment | |
CN117152815A (en) | Student activity accompanying data analysis method, device and equipment | |
CN115512174A (en) | Anchor-frame-free target detection method applying secondary IoU loss function | |
CN114639132A (en) | Feature extraction model processing method, device and equipment in face recognition scene | |
CN111666830A (en) | Intensive crowd counting and detecting frame | |
Wu et al. | Question-driven multiple attention (dqma) model for visual question answer | |
CN110458113A (en) | A kind of non-small face identification method cooperated under scene of face | |
CN117115750B (en) | Application of improved ViT in X-ray security check graph contraband identification | |
Lin et al. | Regulating Balance Degree for More Reasonable Visual Question Answering Benchmark | |
CN117636134B (en) | Panoramic image quality evaluation method and system based on hierarchical moving window attention |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |